Qwen3.8-Flash-Next-DS4-IQ2

Qwen3.8-Flash-Next DS4 Q2 (imatrix)

The default Q2 release is now 44.81 GB (41.73 GiB), including MTP. It uses IQ2_XXS gate/up experts and padded Q2_K down projections, saving 5.16 GiB (11%) over the previous MXFP4-down build with essentially unchanged measured speed. This is the smaller option for memory-constrained Macs, including 64 GB systems.

The external Q4_1 PLE sidecar is required and reused from the Q4 repository. Both files together occupy 76.81 GB (71.53 GiB) on disk. The PLE table is CPU-mapped and demand-paged; resident pages still consume RAM. Start with 8K context and a 1,024-token prefill chunk on a 64 GB Mac. Context buffers, host allocations, and other applications also affect memory use. The measurements below ran on an M3 Ultra with 512 GiB; they do not establish a physical 64 GB fit.

Model structure

• 48 trunk layers, 512 routed experts per layer: the 96 gate/up tensors use imatrix-calibrated IQ2_XXS, unchanged from the previous release.
• 48 trunk down tensors: imatrix-calibrated Q2_K, quantized directly from the original BF16 weights. Each 640-column logical input row is zero-padded to 768 physical columns for three 256-value Q2_K blocks. Metal kernels use the padded weight stride and only the 640 real inputs; the model's architecture and activation width remain unchanged.
• Dense GDN/QSA projections, shared experts and output: Q8_0, preserved from the previous model.
• Token embeddings: BF16. Routers, norms, hyper-connections and control tensors retain their existing F16/F32 precision.
• Embedded MTP block: all 32 tensors preserved, including Q4_K gate/up and MXFP4 down experts. Ordinary decoding and optional --mtp speculation use the same GGUF.
• Required PLE n-gram sidecar: Q4_1, approximately 32.00 GB / 29.80 GiB on disk, shared with the Q4 release. It is not included in the main GGUF.

Only 48 tensor payloads changed. The other 1,207 tensors are byte-identical to the previous release, and all 1,255 output tensors were read back and hash-verified. The .gguf.json file records output hashes, down-weight source hashes, calibration provenance, and the physical padding geometry.

This is a DS4-targeted qwen4exp GGUF. It requires the padded-down runtime support in commit 5bd8796 or later on the qwen3.8-flash-next branch. Older builds reject the new down-tensor layout. Compatibility with other runners is not claimed.

Download and run

Update the DS4 checkout to the current qwen3.8-flash-next branch, then build with make -j8 ds4 ds4-server. Run from that checkout's root:

./download_model.sh qwen38-q2
./ds4 --metal \
  --ple gguf/Qwen3.8-Flash-Next-PLE-Q4_1.gguf \
  --ctx 8192 --prefill-chunk 1024

The script downloads this model and the PLE sidecar, then updates ds4flash.gguf to the new main model. For a custom directory, use:

QWEN_Q2_DIR="$HOME/models/qwen38-ds4-iq2"
hf download ivanfioravanti/Qwen3.8-Flash-Next-DS4-IQ2 \
  Qwen3.8-Flash-Next-IQ2XXSImatrix-Q2KDownPad768-MTP.gguf \
  --local-dir "$QWEN_Q2_DIR"
hf download ivanfioravanti/Qwen3.8-Flash-Next-DS4-Q4 \
  Qwen3.8-Flash-Next-PLE-Q4_1.gguf --local-dir "$QWEN_Q2_DIR"

./ds4 --metal \
  -m "$QWEN_Q2_DIR/Qwen3.8-Flash-Next-IQ2XXSImatrix-Q2KDownPad768-MTP.gguf" \
  --ple "$QWEN_Q2_DIR/Qwen3.8-Flash-Next-PLE-Q4_1.gguf" \
  --ctx 8192 --prefill-chunk 1024

An existing Q4_1 PLE sidecar can be reused without downloading it again. Add --mtp for speculation; use --mtp-exact-sampling when preserving the target sampling distribution is required. The older main-model filename has been replaced, so update any explicit -m paths.

For optional vision, run ./download_model.sh qwen38-vision and add --vision gguf/mmproj-Qwen3.8-Flash-Next-Q8_0.gguf. The encoder is separate and unchanged. Earlier vision smoke scores belong to the superseded model; the tests below cover text, kernel correctness, and a short MTP check.

Measured tradeoff

On 99 BF16-reference continuations / 2,376 tokens, the old MXFP4-down model versus this Q2_K-down release measured:

• BF16 top-token agreement: 91.33% → 90.32%.
• Target NLL: 0.290735 → 0.303570 (4.4% higher; lower is better).
• First-token matches: 69/99 → 65/99.

On the matched 12-question hard-smoke suite, with a 2,048-token answer budget, temperature 0, seed 123, 32K context and no retries or MTP:

• Previous release: 6 correct, 0 wrong, 6 incomplete.
• Current release: 5 correct, 0 wrong, 7 incomplete.

Incomplete means no gradeable final answer. These limited tests show extra quantization drift and do not establish equal reasoning quality.

Two benchmark passes in opposite model order used identical prompts, 1,024-token prefill chunks and 128 teacher-forced decode tokens per frontier:

• 4K: decode 47.63 → 47.31 tok/s; prefill 962.72 → 973.17 tok/s.
• 32K: decode 47.09 → 46.94 tok/s; prefill 926.06 → 935.95 tok/s.

The 32K prefill processes the additional 28,672 tokens after 4K. A separate 128-token greedy smoke test produced identical output with and without MTP. Kernel tests cover padded decode and prefill, and the one-token CPU/GPU check agrees on top-1. See full results and machine-readable measurements.

Provenance and integrity

Source: Qwen/Qwen3.8-Flash-Next, BF16 revision de4b8e4d43b917e7706784d8bb445c9af86a3540. Calibration uses the pinned Unsloth imatrix, revision c8b5954a88c2775c546b92593eda40ea041d3176, SHA-256 a5863123db1ca458727e738955bef7bfc199520aa2bee3a30142a1aff9254154. It contains 45 chunks of 18,432 tokens. Sixteen gate/up and eight down entries with zero activation counts use the deterministic per-expert weight-energy fallback; the remaining entries use measured activation importance.

Build tool and recipe: qwen4_iq2.py --projection down.

Main GGUF SHA-256: 341c8d79468384a05e22998ae834a489145f6c385e276dcf79170a6b0d0ffd2d. SHA256SUMS covers the published release files except itself and the Hub's .gitattributes. The previous weights and their text/vision reports remain available in repository history.

Derived from Qwen's original model, with calibration from Unsloth and the DS4 runtime/build tools. This derivative carries the Qwen Community License 1.0; the original copyright and full terms are in LICENSE.